Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

WATER RESOURCES BULLETIN

FEBRUARY 1992

AMERICAN WATER RESOURCES ASSOCIATION

VOL. 28, NO.1

RISK OF EXTREME EVENTS IN A MULTIOBJECTIVE FRAMEWORK'


Yacov Y Haimes, James H. Lambert, and Duan Li2

their averages, not their extremes, continues to dominate the analyses done by most water resources planners and managers. When the expected value is used

ABSTRACT: This paper summarizes advances made in risk-based


decisionmaking in water resources through use of the partitioned
multiobjective risk method (PMRM). (Risk is a measure of the probability and severity of adverse effects.) In the PMRM, the risk of
extreme events is differentiated from risk involving less extreme
damage severity and is evaluated within a multiobjective framework. Study of the extreme-event risk function f4() has addressed
the following issues: methods for calculating f4(); the sensitivity of
f4(.) to various parameters, particularly to the partitioning point of

as the sole criterion in the analysis of such random


variables, it can lead to fallacious implications for policy formulation. A wide gap exists between what the
expected value of a random variable implies and what
water resources planners and managers intend to use
it for; recognition of this gap' has led to the develop-

the extreme-event range and the selection of probability density


functions; insight provided by the statistics of extremes; and the

ment of the partitioned multiobjective risk method


(PMRM) (Asbeck and Haimes, 1984). Extensions of
the PMRM have been reported in numerous theses
and articles since 1984. This paper provides a brief
guided tour to some of the important results docu-

impact of f4( s) on risk management, for example, in the application

of the PMRM to water resources problems. In particular, this paper


shares with the reader recent research results on the PMRM, the
relationship between the statistics of extremes and the conditional
expected value, derived formulas for f4( s), distribution-free estimates of f4('), documented case studies in dam safety, and future
research directions.
(KEY TERMS: multiobjective decisionmaking; partitioned multiobjective risk method (PMRM); risk assessment; risk management;
statistics of extremes.)

mented in these publications.


To show the progress made on developing an alter-

native measure of risk, focusing on the risk of

extreme events, we briefly review the statistics of


extremes, on which a methodology that is an extension of the PMRM is based. We also summarize applications of the PMRM and its extensions to areas that

INTRODUCTION

have varied from dam safety to hazardous waste to


the Space Shuttle.

If water resources planners and managers were to


contend only with the average precipitation or the

The motivation for preparing this paper is twofold:

average streamfiow, flood, or drought, then the use of


the expected value of such random events would be
adequate for all planning and management activities.
Such a utopian reality would render their work ineffective and unchallenging. Everyone knows, however,
that precipitation, streamfiow, floods, and droughts
assume a very wide range of values in most regions,
and that it is their variabilities and extremes that are
at the heart of planning and management complexity.
Yet, the expected value of these natural events, i.e.,

(1) Risk management inherently leads to a multiobjective analysis, in which trade-offs among costs,
benefits, and risks must be addressed in a multiobjective framework, given the noncommensurate units!
dimensions of each objective function.

(2) There is need for a new family of measures for


risk measures that build on the traditional multiob-

jective analysis and that supplement and complement


the expected-value approach, which has been used as
the sole measure of risk in the past.

'Paper No. 91109 of the Water Resources Bulletin. Discussions are open until October 1, 1992.
2Respectively, Lawrence R. Quarles Professor of Systems Engineering and Civil Engineering and Director, Center for Risk Management of
Engineering Systems, University of Virginia, Olsson Hall, Charlottesville, Virginia 22903; Ph.D. Student, Department of Civil Engineering,
University of Virginia, Charlottesville, Virginia 22903; and Research Assistant Professor, Department of Systems Engineering, University of
Virginia, Charlottesville, Virginia 22903.

201

WATER RESOURCES BULLETIN

Haimes, Lambert, and Li

the expected-value criterion. The optimal action is

Most papers on multiobjective analysis today focus


on the optimization element the generation of noninferior solutions, their associated trade-offs, and the

selected so as to minimize the expected monetary loss,

or, more generally, the expected disutility (Raiffa,


1968; Keeney and Raiffa, 1976). The construction of
the disutility function reflects the decisionmaker's
risk-neutral, risk-prone, or risk-adverse behavior.
Multiattribute disutility functions can be constructed
when more than one system attribute need to be

selection of a preferred solution. Only a few papers,

however, focus on the modeling component the


quantification of the functional relationships among
the building blocks of mathematical models, e.g., random variables, decision variables, state variables, and
objective functions. The following section highlights
the broader context surrounding the PMRM: a quanti-

taken into account. Decisionmaking under uncertain-

ty refers to situations in which the probabilities of


occurrence of the states of nature are unknown.
Several approaches (Sage, 1977) can be adopted in

tative definition of risk, the frameworks of risk


assessment and total risk management, and a comparison of existing methodologies for analysis of risk

such situations in accordance with the decisionmak-

er's attitude toward risk. The Laplace criterion

and uncertainty.

assumes equal probability of occurrence for all states

of nature and selects an optimal action that minimizes the sum of all potential losses. The minimax

METHODOLOQIES FOR DECISIONS


UNDER CONDITIONS OF RISK

criterion is a pessimistic decision rule that minimizes


the maximum potential losses corresponding to differ-

ent actions. In contrast, the mm-mm criterion is an


optimistic decision rule that minimizes the minimum
potential losses corresponding to different actions.
The Hurwicz's rule combines the minimax and mmmm approaches. The optimal action is selected such
that a weighted sum of the least and most favorable
outcomes is minimized. The uncertainty sensitivity

Kaplan and Garrick (1981) define risk as a measure of the probability and severity of adverse effects.
Lowrance (1976) frames the dilemma of decisionmak-

ing under risk in a single epistemological sentence:


"Who should decide on the acceptability of what risks,
for whom, and in what terms, and why?" The process

index method (Haimes and Hall, 1977; Li and

of risk assessment and management is a systemic

Haimes, 1988) selects an action, in a multiobjective


framework, that minimizes the system objective func-

attempt to answer this complex question. Risk management is commonly distinguished from risk assess-

ment, even though some may use the term risk

tion at a nominal value of an uncertain parameter

management to connote the entire process of risk

while minimizing the sensitivity of system output to a


perturbation of the uncertain parameter.
The PMRM builds upon various advantages of the

assessment and management. In risk assessment, the


analyst often attempts to answer the following triplet
set of questions (Kaplan and Garrick, 1981): What
can go wrong? What is the likelihood that it would go
wrong? And, what are the consequences? Answers to
these questions help risk analysts identify, measure,
quantify, and evaluate risks and their consequences

above methodologies. For instance, the PMRM


differentiates among ranges of damage severity in a
spirit similar to those of the minimax and expecteddisutility criteria. While the minimax approach determines a control strategy for the worst-case scenario
without any probabilistic knowledge of the states of
nature, the PMRM provides for a conservative policy

and impacts. Risk management builds on the risk


assessment process by seeking answers to a second
set of three questions (Haimes, 1991): What can be

by using a probabilistic description of the risk of

done? What options are available and what are their


associated trade-offs in terms of all costs, benefits,
and risks? And what are the impacts of current man-

extremes. While an expected disutility may represent


the decisionmaker's strongly risk-averse attitude, the
PMRM avoids the overall commensuration of differ-

ent outcome levels in the disutility function and at


the same time preserves the intuition behind the
expected-disutility criterion for each partitioned risk
region. While the percentile criterion generates a

agement decisions on future options? Only when


these questions are addressed in the broader context
of management, where all options and their associated trade-offs are considered within the hierarchical
organizational structure, can a total risk management

given extremal percentile of a random variable, which


is a single point estimate of loss, the PMRM provides

be realized.

Methodologies for decisionmaking under risk and

an expected-risk measure in the defined range of

Decisionmaking under risk refers to situations in

extremes. Furthermore, the multiobjective representation of risk in the PMRM gives multiple trade-off
information, which contributes to scientifically-sound

uncertainty are abundant in the literature.

which the decisionmaker knows the probabilities of


occurrence of the states of nature. The commonlyused methodology in this case is the Bayes' rule, or
WATER RESOURCES BULLETIN

risk management by accounting for different


202

Risk of Extreme Events in a Multiobjective Framework

perceptions of the potential risk. The PMRM is introduced in the following sections.

TUTORIAL ON THE PMRM

The purpose of this section is to provide a brief


tutorial introduction to the PMRM. A continuous nonnegative random variable X of damages has a cumulative distribution function (CDF) P(x) and a probability
density function (PDF) p(x), which are defined by the
relationships
P(x) = prob[X < xl

0
0

(1)

Damage X

and

p(x)=

Figure la. Probability Density Function.

dP(x)
dx

(2)

The CDF represents the nonexceedance probability of


x. The exceedance probability of x is defined as the
probability that X is observed to be greater than x and
is equal to one minus the CDF evaluated at x. Typical
PDF, CDF, and exceedance probability functions are
plotted in Figures la, ib, and lc, respectively.
The expected value, average, or mean value of the
random variable X is defined as

E[X]= Jx p(x) dx

(3)

In the PMRM, the concept of the expected value of


damage is extended to generate multiple conditional
expected-value functions, each associated with a particular range of exceedance probabilities or their corresponding range of damage seventies. The resulting
conditional expected-value functions, in conjunction
with the traditional expected value, provide a family

Damage X
Figure lb. Cumulative Distribution Function.

of risk measures associated with a particular policy.


Let 1 - a1 and 1 - a2, where 0 < a1 < a2 < 1, denote
exceedance probabilities that partition the domain of

X into three ranges, as follows. On a plot of


exceedance probability, there is a unique damage

on the damage axis that corresponds to the


exceedance probability 1 - a1 on the probability axis.

Similarly, there is a unique damage 2 that corresponds to the exceedance probability 1 - a2. Damages
less than are considered to be of low severity, and

damages greater than 2 are of high severity.


Similarly, damages of a magnitude between and 2

are considered to be of moderate severity. The partitioning of risk into three severity ranges is illustrated

Damage X
Figure ic. Exceedance Probability Function.

203

WATER RESOURCES BULLETIN

Haimes, Lambert, and Li

in Figure 2. If the partitioning probability a1 is specified, for example, to be 0.05, then is the 5th per.centile. Similarly, if 1 - a2 is taken to be equal to 0.05,
then 2 is the 95th percentile.

fx p(x) dx

5p(x) dx
Low severity
High exceedance probability

p1

* )rj

and

Moderate severity
exceedance probability

1 a1

fxpxdx
p2

(6)

fp(x) dx
I

High seventy

p2

Thus, for a particular policy option, there are three


measures of risk, f2('), f3('), and f4(), in addition to
the traditional expected value denoted by f5('). The
function f1(') is reserved for the cost associated with
the management of risk. Note that

1-a2
0

Damage X

Figure 2. Mapping of the Probability Partitioning


Onto the Damage Axis.

fx p(x) dx

f 5()=

f p(x) dx

For each of the three ranges, the conditional

expected damage (given that the damage is within


that particular range) provides a measure of the risk

associated with the range. These measures are

obtained through the definition of the conditional


expected value. Consequently, the new measures of
risk are: f2('), of high exceedance probability and low
severity; f3('), of medium exceedance probability and
moderate severity; and f4('), of low exceedance proba-

bility and high severity. The function f2(') is the


expected value of X, given that x is less than or equal
to f3:

f2() = E[Xjx f3]

since the probability of the sample space of X is necessarily equal to one. In the PMRM, all or some subset

of these five measures are balanced in a multiobjective formulation. The details are made more explicit
in the next two sections.

Assume that the damage severity associated with


the particular policy s, jE (1,... ,q) can be represented
by a continuous random variable X, where px(x;s)
and Px(x;s1) denote the PDF and the CDF of damage,

()

respectively. Two partitioning probabilities, aj, i = 1,2,

fp(x) dx

are preset for the analysis and determine three

ranges of damage severity for each policy sj. A unique


damage, f3jj corresponding to the exceedance probabil-

Similarly, for the other two risk functions f3() and


f4('),
WATER RESOURCES BULLETIN

(7)

GENERAL FORMULATION OF THE PMRM

Jxp(x) dx

= fx p(x) dx

ity (1 - aj), can be found due to the monotonicity of


Px(x;s). The policies s, the partitions a1, and the
204

Risk of Extreme Events in a Multiobjective Framework

bounds
pression

If the unconditional expected value of the damage

of damage ranges are related by the ex-

from policy s is defined to be f5(s3) and the denomina-

tor of Equation (10) is defined to be 0, i = 2,3,4, the


i = 1,2

Px(I3;s3) = ai

following relationship holds:

(8)

This partitioning scheme is illustrated in Figure 3 for

two hypothetical policies s1 and s2. The ranges of


damage severity include high exceedance probability
and low damage, (X: x e [Oj'1j]) the set of possible
realizations of X for which it is true that x e [3oJi];
medium exceedance probability and medium damage,
(X: x E [f31J,12J]); and low exceedance probability and

with the O 0 and 2 + 03 + 04 = 1. The O are the


probabilities that X is realized in each of the three
The preceding discussion has described the parti-

tioning of three damage ranges by selected

exceedance probabilities ct, i = 1,2. Alternatively, the


PMRM provides for the partitioning of damage ranges

by preset thresholds of damage. For example, the


meaning of f4(s) in partitioning by a fixed damage

x.

The conditional expected-value risk functions f1, i =

becomes the expected damage resulting from policy j


given that the damage exceeds a fixed magnitude. For
further details on the partitioning of damage ranges,

2,3,4, are given by

fj(s) = E(X I px(x;s3), x E [i-2ji-1jI)

see Asbeck and Haimes (1984) and Karlsson and


Haimes (1988a, b).

(9)

j = 1,...,q

(11)

damage ranges and are independent of the policies Sj.

high damage (extreme event), [X: x e [132j,fl3j]), where


fOj and 13j are the lower and upper bounds of damage

i = 2,3,4;

f5(s) = 02f2(s3) + 93f3(s3) + O4f4(s)

The conditional expected-value functions in the


PMRM are multiple, noncommensurate measures of
risk, each associated with a particular range of dam-

and, equivalently,

age severity. In contrast, the traditional expected

value commensurates risks from all ranges of damage


severity and represents only the central tendency of
the damage. A risk-cost multiobjective optimization
problem utilizing the conditional expected-value functions is formalized in the next section.

i 1.j

Jx

f.(s.)=

2j

i=2,3,4; j=1,...,q

lli1.j

pxj',c

$
I 2,j

(10)

1-a1

'-4

o
0

21

I22

Damage X

Figure 3. Mapping of the Probability Partitioning Onto the Damage Axis for Two Policies.

205

WATER RESOURCES BULLETIN

Haimes, Lambert, and Li

transfers the virtues of the (unconditional) expected


value to a preselected extreme range of damage, thus

MULTIOBJECTIVE OPTIMIZATION
IN THE PMRM

providing a basis for conservative policies.


If the cost of policy sj is denoted by f1(s), a family of
multiobjective optimization problems can be defined

as

mm

f 1Cs.)
j

si

f.(s.)

Ii

i=2,3,4,5

(12)
S2

in which the traditional two-objective optimization


approach
mm

f 1Cs.)

f5(s)

Si

(13)

f4

is augmented by risk-cost optimization involving a


family of risk functions corresponding to various
ranges of damage severity. Some or all of the conditional expected-value functions, in conjunction with

f5(s), f4(s)
Figure 4. Trade-Offs Between Cost and Risk.

f5(sj), may be seen as representative of the sj in a par-

ticular decisionmaking context. Figure 4 illustrates


that the marginal cost between the two policies s1 and

2 can be associated with both a small increment of


the expected value f5(s) and a large increment of
extreme-event risk f4(sj). In this case, the function
f4(sj) conveys important additional information that
would be hidden by the use of f5(s) alone. Solving the

RECENT DEVELOPMENTS IN THE STUDY


OF EXTREME-EVENT RISK

The study of f4(') has focused on the following


questions, all of which are aimed at making f4()

multiobjective problems posed by Equation (12),

more amenable to its use in the management of risk:

depends on the analyst's interaction with the decisionmaker in determining an optimal balance of risk

able for calculating f4( )?

which is the domain of multiobjective decisionmaking,

(1) What efficient formulas or methods are avail-

(2) How sensitive is f4(') to various modeling

functions and cost. In other words, locating this optimal balance is equivalent to determining the level of
safety, where safety is defined as the level of risk that
is deemed acceptable to the decisionmaker (Lowrance,
1976). The surrogate worth trade-off (SWT) method
(Haimes and Hall, 1974) can be used to find a pre-

parameters? What is the sensitivity of f4(') to the partitioning point of the range of extreme events?

(3) What is the contribution of the statistics of


extremes to the understanding of f4(')?

(4) What is the impact of f4(') on risk management, for example, in the applications of the PMRM
to water resources problems?

ferred policy from among noninferior, or Paretooptimal, policies.

The risk function that has been studied most is the


low-exceedance-probability, high-damage function,
f4('). As stated in Equations (9) and (10), f4(') is the
(conditional) expected value of damage given that the
damage is realized in the most severe, extreme-event

The sections that follow discuss progress that has


been made in these directions.
Statistics of Extremes

range. For example, the extreme-event risk, f4('),


could be the expected flood damage given that the
flood magnitude is in the worst (largest) 10 percent

The subject of the statistics of extremes is concerned with studying the largest (or smallest) value
realized from a sample of t independent, identically

range (equivalently, the flood magnitude exceeds the


90th percentile, or a2 = 0.9).
The function f4() is a natural measure of the risk
of extreme events for risk-based decisionmaking. It
WATER RESOURCES BULLETIN

distributed (i.i.d.) random variables (Ang and Tang,


1984). The study of f4() has been advanced by the
206

Risk of Extreme Events in a Multiobjective Framework

correspondence between the function f4('), the statistics of extremes, and hydrologic time series phenomena (Karlsson and Haimes, 1988a, b).

The statistics-of-extremes approach additionally


gives important results concerning the sensitivity of

f4(') to the range (partitioning point) of extreme

Just as statistical moments parametrize random


variables, the statistics of extremes defines two
parameters that characterize the largest value of a
sample of t i.i.d. variables, which is itself a random

events and the relationships between f4() and the

asymptotic nature of the damage distribution


(Karlsson and Haimes, 1988a, b). Distribution-free
estimates of f4(') that were obtained by using the
statistics of extremes are discussed in a separate seetion.

variable. The characteristic largest value, ut(sj), is of


the magnitude that is exceeded on the average once in
t realizations of X, and it is implicitly defined by
t[lPx(ut(s)] =

Exact Formulas: Normal, Lognormal, and Weibull

(14)

Distributions of Damage

A second parameter, (sj), which measures the sensitivity of the characteristic largest value, ut(sj), to the

Analytical, closed-form expressions for f4() in


terms of well-known mathematical functions (error
and gamma functions) have been derived for cases in
which the random variable representing damage is of

sample size t, is defined by

du t(s?

d[ln(t)]

(15)

(s)

normal,' lognormal, or Weibull distributions (Romei et

al., 1992). The results facilitate computation of f4(')


and establish functional relationships concerning the
sensitivity of f4() to the partitioning point of extreme
events and to the parameters of the damage distribution.
For instance, if the damage distribution for policy
s is normal with mean and standard deviation a3,

From Equation (14) it can be seen that if X is an


annual maximum flood, then t is the return period in

years of the event of magnitude ut(sj). The return


period corresponding to the partition probability a2 of
extreme events is given by

tien

(16)

(1- a2)

f4(2;

From Equations (8), (14), and (16), it can be seen that


the damage partition 12j for extreme events is exactly
the characteristic largest value, Ut(S). Equation (10),
with i = 4, is rewritten via the statistics of extremes

a.)= .+ af*[2u3]

(19)

where f4*(.) is the f4 function for the standard normal


variate and is defined by Romei et al. (1992), in terms
of the error function. This exact result has advantages

as

over statistics-of-extremes approaches that use


approximations. Equation (19) provides an exact corn-

4j

f (s .) = t f-i- u(s .) d'c

putational tool when the damage is normal, and its


derivatives with respect to the partition point f2j and
the parameters jij and a3 generate illuminating sensitivity results. Exact denvations of f4(') and its sensitivities are also demonstrated for the lognormal and

(17)

It can be derived from Equation (17) that f4(sj) can be


approximated, with an appropriate number of terms,
through the expression

Weibull distributions (Romei et al., 1992).

d u(s.)
f4(s.)=u(s.)+

k=1 d(ln t)

Distribution-Free Results

+c(s.)t

(18)

In risk-based decisionmaking under uncertainty,


the type (e.g., normal, lognormal, Weibull, etc.) of
damage distribution may be unknown. In these cases,
a risk measure that is independent of the knowledge

where the term c(sj) is typically equal to zero. Equation (18), apart from being theoretically interesting, is
particularly useful for engineering models in which
monetary or other damage is a transformation of a
hydrologic random variable.

of the underlying distribution type can be used


(Mitsiopoulis and Haimes, 1989). Through the
Chebyshev inequality, an upper bound on the function

207

WATER RESOURCES BULLETIN

Haimes, Lambert, and Li

f4(') and the upper bound of its sensitivity to the partitioning point are generated.
Independent of any particular distribution of dam-

Distributions that are asymptotically of Type II

age, with statistical moments and YJ, it can be

of f4() for random variables that are asymptotically

shown that

of Type III is given by

f4(s.) min[i2. J.t.+ 2a.'V't]

include the lognormal, Pareto, Cauchy, and others.


The corresponding distribution-free approximation

f4()=u+

(20)

where w is the upper limit of the damage random


variable. Distributions that are asymptotically of

when t is greater than the return period of the mean,

This bound can be used as a surrogate or estimate


of f4(s3) when the damage distribution is uncertain. A

Type III include the uniform, the triangular, and most


distributions that have a finite upper bound.

distribution-free bound on the sensitivity to the

The extremal parameters ut and 8 of damage in


many situations can be estimated using either the
method of moments or extremal probability paper

extreme-event partitioning point is given by

df (s.)
dt

t. 3

1] (24)

(Mg and Tang, 1984). These new approximations are


very powerful in that they absolve the analyst of the
necessity of assuming a particular probability distribution of damage, a step that is often most uncertain
in practice.

(21)
.1

An appreciation of the sensitivity of f4(sj) to the partition probability is a necessary step toward the robust
assessment of extreme-event risk.

APPLICATIONS OF THE PMRM

Approximating Catastrophic Risk Through the


Statistics of Extremes

Application of the PMRM to dam safety has shown

that f4(') conveys additional information about risk


when used together with the expected value of dam-

Distribution-free estimates of f4(') have been


derived for each of three extreme-value types, Gumbel

age (Karlsson and Haimes, 1989; Haimes et al., 1988;

'I'ypes I, II, and III, based only on knowledge of the


two extremal parameters, ut and 6. If a random variable of damage can be assumed to be asymptotically

Petrakian et al., 1989). Understanding of the potential extreme consequences of a remedial action or dam

design is enhanced by the risk measure f4(') of

of one of the three extreme-value types, then the

extreme events. In case studies cited in this section,


the cost of flood damage (flood risk) is modeled as a
function of flood magnitude, which has stochastic
variability, for different dam modification actions. The
relationship between the risk function f4(') and the
magnitude and return period of the probable maxi-

approximations generate distribution-free estimates


of the extreme-event risk f4(') that rely solely on the
values of the statistics-of-extremes parameters Ut and
6 (Mitsiopoulis et al., 1991). These approximations
based on the statistics of extremes are nearly exact
for large values of t, and, equivalently (see Equation

mum flood (PMF) is explored by Lambert et al.

16), large values of the partition probability a2.


For a random variable of damage that is asymptotically of Gumbel Type I, the approximation to f4() is
given by

f4()=ut+ 1it

(1992a). Additionally, the sensitivities of f4() to the


flood distribution (normal, lognormal, etc.) and to the
form of the stage-damage function have been analyzed by Karlsson and Haimes (1989). Distribution-

free approximations, based on the statistics of

(22)

extremes, to f4(') are being studies for use when the


underlying distribution of flood is uncertain (Lambert

Distributions that are asymptotically (largest values)


gamma, Rayleigh, exponential, and others.

et al., 1992a).
In risk management of large-scale systems, such as
spacecraft and hazardous waste sites, it is often bene-

Type II, the approximation to f4() is given by

ficial to prioritize many components of a system

of Type I include the normal, Gumbel, Weibull,


For a random variable that is asymptotically of

according to their individual contributions to the overall system's risk. Adoption of the PMRM for the ranking of components facilitates the differentiation of the
risks of extreme- and of moderate-severity events in a

f4()=u+ l/+(l/) /(u

(23)

multiattribute ranking approach. Elements of the


WATER RESOURCES BULLETIN

208

Risk of Extreme Events in a Multiobjective Framework

PMRM were incorporated within a methodology


developed for ranking critical components of the

Kaplan, S. and B. J. Garrick, 1981. On the Quantitative Definition


of Risk. Risk Analysis 1(1):11-27.
Karlsson, P.-O. and Y. Y. Haimes, 1988a. Probability Distribu-tions

Space Shuttle (Haimes et al., 1991). The methodology

and Their Partitioning. Water Resources Research 24(1):

was further extended for use in the ranking of hazardous waste sites for their subsequent remediation

21-29.

Karisson, P.-O. and Y. Y. Haimes, 1988b. Risk-Based Analysis of


Extreme Events. Water Resources Research 24(1):9-20.

(Lambert et al., 1992b).

Karisson, P.-O. and Y. Y. Haimes, 1989. Risk Assessment of


Extreme Events: An Application. Journal of Water Resources
Planning and Management 115(3):299-320.
Keeney, R. L. and H. Raiffa, 1976. Decisions With Multiple Objec-

tives: Preferences and Value Tradeoffs. John Wiley and Sons,

FURTHER INVESTIGATIONS

New York, New York.

Lambert, J. H., Y. Y. Haimes, T. Dolezal, and D. Li, 1992a. Risk of

It is hoped that the theoretical and methodological


results presented here entice researchers and practi-

Extreme Flood Damage: Sensitivity to Return Period of PMF


(manuscript).

Lambert, J. H., C. W. Ling, and Y. Y. Haimes, 1992b. Remediation

tioners alike with the challenge of exploring this

Site Prioritization by the Risk Ranking and Filtering Method.

almost terra incognita of the risk of extreme events.


Many policies and procedures have been formulated
on the basis of the centrality of the expected value.
Their soundness and representativeness should be

In: Risk-Based Decision Making in Water Resources V,


Proceedings of an Engineering Foundation Conference, Y. Y.
Haimes, D. A. Moser, and E. Z. Stakhiv (Editors). American
Society of Civil Engineers, New York, New York (in press).

Li, D. and Y. Y. Haimes, 1988. The Uncertainty Sensitivity Index


Method (USIM) and Its Extension. Naval Research Logistics

questioned and the associated analyses should be


revisited. Multivariate probability density functions

35:655-672.

Lowrance, W. W., 1976. Of Acceptable Risk: Science and the


Determination of Safety. William Kaufmann, Inc., Los Altos,

offer unique research opportunities to scholars interested in the risk of extreme events, where the use of
decomposition and hierarchical coordination applied
to statistics of extremes provides one avenue of tackling complex problems.

California.

Mitsiopoulis, J. and Y. Y. Haimes, 1989. Generalized Quantification of Risk Associated With Extreme Events. Risk analysis
9(2):243-254.

Mitsiopoulis, J., Y. Y. Haimes, and D. Li, 1991. Approxi-mating

Catastrophic Risk Through Statistics of Extremes. Water

Resources Research 27(6):1223-1230.


Petrakian, R., Y. Y. Haimes, E. Stakhiv, and D. A. Moser, 1989. Risk

ACKNOWLEDGMENTS

Analysis of Dam Failure and Extreme Floods. In: Risk Analysis

The research summarized in this paper was supported in part


by the National Science Foundation under Grant BCS-8912630,
entitled "Integrating the Statistics of Extremes With Conditional

and Management of Natural and Man-Made Hazards, Y. Y.


Haimes and E. Stakhiv (Editors). American Society of Civil

Expectation." Editorial assistance was provided by Virginia

Raiffa, H., 1968. Decision Analysis: Introductory Lectures on

Engineers, New York, pp. 8 1-122.

Choices Under Uncertainty. Random House, New York, New

Benade.

York.

Romei, S. F., Y. Y. Haimes, and D. Li, 1992. Exact Deter-mination


and Sensitivity Analysis of a Risk Measure of Extreme Events.
Information and Decision Technologies (in press).
Sage, A. P., 1977. Methodology for Large Scale Systems. McGraw-

LiTERATURE CiTED

Ang, A. H-S. and W. H. Tang, 1984. Probability Concepts in

Hill, New York, New York.

Engineering Planning and Design, Volume II: Decision, Risk,


and Reliability. John Wiley and Sons, New York, New York.
Asbeck, E. L. and Y. Y. Haimes, 1984. The Partitioned Multiobjective Risk Method (PMRM). Large Scale Systems 6(1):13-38.

Haimes, Y. Y., 1991. Total Risk Management. Risk Analysis


1i(2):169-171.

Haimes, Y. Y. and W. A. Hall, 1974. Multiobjectives in Water


Resources Systems Analysis: The Surrogate Worth Trade-off
Method. Water Resources Research 10:615-624.
Haimes, Y. Y. and W. A. Hail, 1977. Sensitivity, Responsivity, Sta-

bility, Irreversibility as Multiple Objectives in Civil Systems.


Advances in Water Resources 1(2).

Haimes, Y. Y., D. Li, J. Pet-Edwards, J. Lambert, V. Thlsiani, and


D. Tynes, 1991. Ranking of Space Shuttle FMEA/CIL Items: The
Risk Ranking and Filtering (RRF) Method. Technical report sub-

mitted to the Vitro Corporation and NASA Headquarters,

Center for Risk Management of Engineering Systems,


University of Virginia, Charlottesville, Virginia.

Haimes, Y. Y., R. Petrakian, P.-O. Karisson, and J. Mitsiopoulis,


1988. Multiobjective Risk-Partitioning: An Application to Dam
Safety Risk Analysis. 1WR Report 88-R-4, Institute for Water
Resources, U.S. Army Corps of Engineers, Fort Belvoir, Virginia.

209

WATER RESOURCES BULLETIN

You might also like